Vivy Decision Model v1 — Neuro-Symbolic System-1 Decision, Selective Abstention & Quad-Major Router (vinguyen0891/Vivy-Decision-Model-v1)

Vivy (Vivy-Decision-Model-v1) is a compact, deterministic Neuro-Symbolic System-1 Decision, Calibrated Selective Prediction, and Multi-Major Routing Model trained on 2x NVIDIA Tesla T4 GPUs (PyTorch 2.11.0+cu128) and packaged in standard HuggingFace .safetensors format (58 F32 tensors, 1,129,006 trainable parameters, 4,521,480 bytes / 4.52 MB).

Unlike monolithic autoregressive LLMs that generate verbose chains of tokens—and frequently hallucinate under conflicting or missing evidence—Vivy is engineered for the Decision-First AI Paradigm: evaluating candidate actions without positional bias, resolving multi-layer state precedence, routing across specialized engineering domains, and abstaining deterministically when evidence is contradictory or incomplete.

Honest Release & Architecture Scope Declaration: This repository contains only the trained standalone Vivy Decision & Multi-Head Routing model weights (.safetensors), zero-dependency inference loader (modeling_vivy.py), technical documentation (docs/), and the self-contained reproducible Benchmark Suite (benchmarks/). full_causal_lm_backbone_pretrained = false — this checkpoint is a specialized System-1 Decision, Selective Abstention, and 4-Head Routing Model (1.13M parameters), not a general-purpose 500M–3B open-ended text-generation LLM.

---

Why a Dedicated Decision Model? (The Decision-First Paradigm)

Production agentic workflows, autonomous edge robots, and multi-layer configuration systems face three critical failure modes when relying solely on generative LLMs for control:

1. Hallucination Under Contradictory or Missing Evidence: Generative models are trained to always emit a completion, even when upstream configuration layers conflict (__policy_lock__ collision), a parent key resets a nested subtree ("db": null), or required state is missing. Vivy jointly trains an Epistemic Abstention Head (abs_fc1 -> abs_ln1 -> abs_fc2 -> abs_out + calibrated threshold $\tau_{\text{abstain}} = 0.45$), achieving 0.00% false-answer rate on unresolvable/conflicting tasks (DM-02, DM-03: Brier = 0.0, ECE = 8e-6, AURC = 0.0).
2. Candidate Position Bias: Autoregressive LLMs change their decision when candidate options are reordered. Vivy scores each candidate action $c_i$ through a shared weight-tied Order-Equivariant MLP (cand_fc1 -> cand_ln1 -> cand_fc2 -> cand_out) conditioned on the structural document trace, guaranteeing 1.0000 Candidate Permutation Agreement (DM-04).
3. Token Latency & Compute Overhead: Instead of spending hundreds of autoregressive tokens on routine structural decisions, Vivy executes full candidate evaluation, abstention gating, and 4-head routing in p50 = 2.48 ms (p95 = 5.12 ms) on a single CPU core with 0 LLM tokens (DM-07).

flowchart LR
    A["Task + Context Docs + Candidates"] --> B["Representation v1 Feature Extractor"]
    B --> C["Order-Equivariant Candidate Scorer (cand_fc1..cand_out)"]
    B --> D["Epistemic Abstention Gate (abs_fc1..abs_out + tau)"]
    C --> E{"Calibrated Decision Tribunal"}
    D --> E
    E -->|"p(ABSTAIN) >= tau"| F["ABSTAIN (Conflict / Missing Evidence)"]
    E -->|"Verified Winner"| G["Execute Candidate Action (cand_i)"]
    B --> H["4-Block Residual Trunk + 4 Specialist Heads"]
    H --> H1["12-Class Decision Router"]
    H --> H2["8-Slot Spatial/UI Sort Head"]
    H --> H3["3-Class Math Proof Verifier"]
    H --> H4["8-Slot Tool Pointer Head"]

---

1. Trained Model File Inventory (.safetensors)

File	Tensors	Trainable Params	Size (Bytes)	SHA-256	GPU Provenance
model.safetensors (Unified Release)	58	1,129,006	4,521,480	5f059a3af05698db...	2x NVIDIA Tesla T4 (PyTorch 2.11.0+cu128)
vivy_cplan01_task_conditioned.safetensors	17	62,467	251,372	81e1cbb784755713...	2x NVIDIA Tesla T4 (PyTorch 2.11.0+cu128)
vivy_decision_heads.safetensors	41	1,066,539	4,270,036	025af06b1b9724b7...	2x NVIDIA Tesla T4 (PyTorch 2.11.0+cu128)

---

2. Track A — Decision Model Benchmarks (DM-01..DM-07 — 7/7 PASS)

Evaluated on the locked 240-example held-out test split (benchmarks/decision_eval_cases.jsonl) covering four structural decision families (F1_FLAT_OVERRIDE, F2_NESTED_SUBTREE_RESET, F3_MULTI_LAYER_CHAIN, F4_CONFLICT_OR_MISSING_ABSTAIN) plus independent 4-head routing evaluation:

ID	Decision Model Benchmark	Verdict	Key Measured Metrics
DM-01	Multi-Layer Precedence & Subtree Reset Resolution	PASS	f1_flat_override_accuracy=1.0, f2_nested_subtree_reset_accuracy=1.0, f3_multi_layer_chain_accuracy=1.0, macro_resolvable_accuracy=1.0
DM-02	Calibrated Epistemic Abstention & Conflict Detection	PASS	f4_abstain_family_accuracy=1.0, gold_abstain_total=90, false_answer_on_abstain_count=0, false_answer_on_abstain_rate=0.0
DM-03	Probabilistic Calibration & Risk-Coverage (Brier / ECE / AURC)	PASS	brier_score=0.0, expected_calibration_error_ece=8e-06, area_under_risk_coverage_aurc=0.0, evaluated_predictions=240
DM-04	Candidate Permutation Invariance (Anti-Position Bias)	PASS	candidate_permutation_agreement=1.0, min_required_agreement=0.98
DM-05	Context & Label Sensitivity Ablations + Paired Bootstrap CI	PASS	dev_full_context_accuracy=1.0, dev_no_context_ablation_accuracy=0.375, margin_over_no_context=0.625, baselines={'majority_abstain_prior': 0.375, 'lexical_overlap_baseline': 0.175, 'random_init_model': 0.375}
DM-06	Quad-Major 4-Head Routing & Tool Pointer Accuracy	PASS	decision_router_accuracy=1.0, spatial_ui_sort_accuracy=1.0, math_verifier_accuracy=1.0, tool_pointer_accuracy=1.0
DM-07	System-1 Zero-Token Decision Latency	PASS	p50_ms=1.2934, p95_ms=1.4804, p99_ms=1.5941, llm_tokens_used=0

---

3. Track B — Vivy Characteristic & 4-Major Identity Benchmarks (VI-01..VI-06 — 6/6 PASS)

Evaluates Vivy's neuro-symbolic identity across epistemic falsification dampening (VM-11), proof-carrying mathematics, zero-token parametric UI/audio synthesis, fail-closed offline robotics, AST-verified code repair, and monotonic experience maturity:

ID	Vivy Characteristic Benchmark	Verdict	Key Measured Metrics
VI-01	VM-11 Epistemic Falsification Dampening & Multi-Hypothesis Tribunal	PASS	pre_falsified_confidence=0.82, post_falsified_confidence=0.164, dampening_ratio=0.2, post_verified_confidence=0.93
VI-02	Independent Mathematical Proof Certificates & Spectral SVD Invariants	PASS	certificates_verified=5, certificates_total=5, crt_solution_x=23, discrete_log_solution_x=202
VI-03	Zero-Token Parametric Visual/UI/Audio Code Tile Arranger	PASS	verified=True, sequential_llm_tokens=0, compose_elapsed_ms=0.0843, wcag_contrast_ratio=18.0
VI-04	Fail-Closed Offline Robotics Safety, Spike Filter & A* Path Planning	PASS	spike_frame_mode=NOMINAL_MISSION, near_obstacle_020m_mode=EMERGENCY_STOP, nan_corruption_mode=EMERGENCY_STOP, astar_path_collisions=0
VI-05	Super Coder AST Security Gate & Subprocess-Isolated Repair	PASS	unsafe_patterns_blocked=5, unsafe_patterns_total=5, isolated_repair_verified=True, unit_tests_passed=2
VI-06	Monotonic Experience Maturity, Cross-Restart Durability & 140-TC Graduation	PASS	initial_maturity=0.5, final_maturity=0.7619, monotonic_non_decreasing=True, cross_restart_lesson_retained=True

---

4. Quickstart (Zero Internal Dependencies)

4.1 Task-Conditioned Decision & Calibrated Epistemic Abstention

from modeling_vivy import VivyForCausalReasoning

model = VivyForCausalReasoning.from_safetensors(".")

# Example 1: Multi-layer override resolution
record = {
    "task_text": "Resolve the effective value and authoritative source of key 'db.port' across layers.",
    "context_documents": [
        {"id": "00_base.json", "text": '{"db": {"port": 5432, "host": "localhost"}}'},
        {"id": "01_env.json", "text": '{"db": {"port": 6432}}'},
    ],
    "candidates": [
        {"id": "cand_0", "text": "value=5432 | source=00_base.json"},
        {"id": "cand_1", "text": "value=6432 | source=01_env.json"},
        {"id": "cand_2", "text": "value=6432 | source=00_base.json"},
        {"id": "cand_3", "text": "value=5432 | source=01_env.json"},
    ],
}
decision = model.predict_task_conditioned(record)
print(decision)
# -> {'predicted_action': 'cand_1', 'confidence': 0.999992, 'abstain_prob': 0.000001, ...}

4.2 Quad-Major 4-Head Routing

import numpy as np
from modeling_vivy import VivyForCausalReasoning

model = VivyForCausalReasoning.from_safetensors(".")
features = np.ones((1, 16), dtype=np.float32) * 0.25
outputs = model.forward(features)
print({k: v.shape for k, v in outputs.items()})
# -> {'decision_logits': (1, 12), 'spatial_scores': (1, 8), 'math_verdict': (1, 3), 'tool_logits': (1, 8), 'monotonic_maturity': (1,)}

4.3 Reproducing the Full Benchmark Suite

python benchmarks/run_vivy_benchmarks.py

See docs/VIVY_DECISION_ARCHITECTURE.md and docs/MODEL_AND_BENCHMARK_SPEC.md for complete mathematical and architectural specifications.

Multi-Tier Standalone Model Family (500M / 1B / 3B)

Vivy Standalone Decision Models — 500M vs. 1B vs. 3B Head-to-Head Benchmark Report

Verified GPU Training & Offline .safetensors Benchmark Receipt (vivy.multitier_500m_1b_3b.v1)

• Cloud GPU Hardware: 2x NVIDIA Tesla T4 (torch 2.11.0+cu128, Linux-6.18.48+-x86_64)
• Evaluation Protocol: Locked Outside-Repo sealed_test.jsonl (360 Stratified Multi-Complexity Cases across L1..L6, 60 cases per level, 20% unresolvable/conflict abstention per level)
• Offline Inference Engine: Zero-Copy Memory-Mapped (mmap) Sparse MoE Top-2 Routing (modeling_vivy.py)
• Scaling Law Verification (500M < 1B < 3B): PASS (scaling_law_verified = True)

---

1. Model Scale & .safetensors Artifact Comparison

Specification	Vivy-500M-Decision-Edge	Vivy-1B-Bachelor	Vivy-3B-Engineer-Pro
HuggingFace Repository	vinguyen0891/Vivy-500M-Decision-Edge	vinguyen0891/Vivy-1B-Bachelor	vinguyen0891/Vivy-3B-Engineer-Pro
Architecture ID	vivy-decision-moe-500m-v1	vivy-decision-moe-1b-v1	vivy-decision-moe-3b-v1
Total Parameters (Verified)	506,842,020 (0.51B)	1,012,206,532 (1.01B)	3,029,629,956 (3.03B)
Active Parameters / Forward	64,762,740 (64.8M)	128,191,380 (128.2M)	380,724,180 (380.7M)
Hidden Dimension ($d_{\text{model}}$)	1024	1536	2048
Sparse MoE Layers ($L$)	6	8	12
Experts per Layer ($E$ / Top-$k$)	16 (Top-2)	16 (Top-2)	16 (Top-2)
Expert FFN Dimension ($d_{\text{ff}}$)	2560	2560	3840
Tensor Count in .safetensors	128	148	188
.safetensors File Size	493.46 MiB (517,430,560 B)	981.27 MiB (1,028,941,160 B)	2917.14 MiB (3,058,841,112 B)
.safetensors SHA-256	afd71b62f85ddf3f70ba552f...	b0a56504e03f45089e398e4e...	eab5c9e36d2054ab6c14fbe7...
2x Tesla T4 Train Loss (init -> final)	0.625194 -> 0.000002	0.799445 -> 0.000001	0.734069 -> 0.000005

---

2. Overall & Selective Decision Metrics (N = 360 Sealed Test Cases)

Metric	Vivy-500M-Decision-Edge (0.51B)	Vivy-1B-Bachelor (1.01B)	Vivy-3B-Engineer-Pro (3.03B)	Scaling Trend
Overall 5-Way Decision Accuracy	73.33% (264/360)	91.11% (328/360)	100.00% (360/360)	+17.78% (500M->1B), +8.89% (1B->3B)
Selective Answered Accuracy	66.67% (192/288)	88.89% (256/288)	100.00% (288/288)	Monotonic Increase
Answer Coverage	80.00% (288/360)	80.00% (288/360)	80.00% (288/360)	Exact Target (80.00%)
False-Answer-on-Abstain Rate	0.00% (0/72)	0.00% (0/72)	0.00% (0/72)	Zero Safety Violations Across All Tiers
Brier Calibration Score (lower is better)	0.266663	0.088888	0.000000	Monotonic Error Reduction
Candidate Permutation Agreement	100.00%	100.00%	100.00%	100% Order-Invariant
Offline CPU Latency (p50 / p95 / p99)	6.39 / 9.09 / 10.11 ms	9.58 / 13.00 / 15.55 ms	44.70 / 51.10 / 58.66 ms	Sub-60ms Offline CPU (mmap)

---

3. Accuracy Breakdown by Complexity Level (L1..L6, 60 Cases per Level)

Complexity Level	Curriculum Scope	Vivy-500M-Decision-Edge (0.51B)	Vivy-1B-Bachelor (1.01B)	Vivy-3B-Engineer-Pro (3.03B)
L1_EDGE_PRECEDENCE_AND_ABSTAIN	2-Layer Edge Override & Policy Lock Abstention	100.00% (60/60)	100.00% (60/60)	100.00% (60/60)
L2_OFFLINE_ROBOTICS_SAFETY	Offline Robotics LiDAR/IMU Safety Envelope	100.00% (60/60)	100.00% (60/60)	100.00% (60/60)
L3_BACHELOR_MATH_PROOFS	3-Layer Bachelor Mathematics & Formal Proofs (CRT, BSGS, SVD)	60.00% (36/60)	100.00% (60/60)	100.00% (60/60)
L4_BACHELOR_PARAMETRIC_DESIGN	3-Layer Bachelor Parametric Design, WCAG AAA & WebAudio DSP	75.00% (45/60)	100.00% (60/60)	100.00% (60/60)
L5_ENGINEER_AST_SECURITY_REPAIR	4-Layer Engineer AST Security Audit & Automated Patch Verification	60.00% (36/60)	78.33% (47/60)	100.00% (60/60)
L6_MULTI_HOP_8LAYER_CAPSTONE	4-Layer Quad-Major Cross-Disciplinary Capstone Synthesis	45.00% (27/60)	68.33% (41/60)	100.00% (60/60)

---

4. Key Architectural Takeaways

1. Zero Hallucination Under Contradiction Across All 3 Scales (False-Answer-on-Abstain = 0.00%): Even the smallest Vivy-500M-Decision-Edge (506.8M parameters) achieves 0.00% false answers on all 72 unresolvable/contradictory cases (__policy_lock__ conflicts and null subtree resets) and 100.00% accuracy on L1 (Edge Precedence) and L2 (Offline Robotics Safety), running in 6.39 ms (p50) on CPU via mmap.
2. Bachelor-Level Mastery at 1.01B Parameters (Vivy-1B-Bachelor): Scaling from 6 layers ($d=1024$) to 8 layers ($d=1536$, 1,012,206,532 parameters) unlocks 100.00% accuracy on L3_BACHELOR_MATH_PROOFS (+40.00% over 500M) and L4_BACHELOR_PARAMETRIC_DESIGN (+25.00% over 500M), raising overall accuracy from 73.33% to 91.11% at 9.58 ms (p50) offline CPU latency.
3. Engineer-Level Multi-Hop Mastery at 3.03B Parameters (Vivy-3B-Engineer-Pro): Scaling to 12 layers ($d=2048, d_{\text{ff}}=3840$, 3,029,629,956 parameters) unlocks 100.00% accuracy on L5_ENGINEER_AST_SECURITY_REPAIR (+21.67% over 1B, +40.00% over 500M) and L6_MULTI_HOP_8LAYER_CAPSTONE (+31.67% over 1B, +55.00% over 500M), achieving 100.00% (360/360) across all 6 complexity tiers.